Inapproximability of Finding Sparse Vectors in Codes, Subspaces, and Lattices

TL;DR

This paper establishes strong NP-hardness inapproximability results for finding sparse vectors in codes, subspaces, and lattices, using novel techniques that extend previous hardness results to all p-norms.

Contribution

It introduces a new reduction approach that bypasses the PCP theorem, extends inapproximability to all p-norms, and simplifies the proof of hardness for the minimum distance problem.

Findings

NP-hardness of approximating sparsest vectors within any constant factor

Extension of hardness results to all p ≥ 0 norms

Elementary derandomization of the reduction for finite fields

Abstract

Finding sparse vectors is a fundamental problem that arises in several contexts including codes, subspaces, and lattices. In this work, we prove strong inapproximability results for all these variants using a novel approach that even bypasses the PCP theorem. Our main result is that it is NP-hard (under randomized reductions) to approximate the sparsest vector in a real subspace within any constant factor; the gap can be further amplified using tensoring. Our reduction has the property that there is a Boolean solution in the completeness case. As a corollary, this immediately recovers the state-of-the-art inapproximability factors for the shortest vector problem (SVP) on lattices. Our proof extends the range of $\ell_p$ (quasi) norms for which hardness was previously known, from $p\geq 1$ to all $p\geq 0$, answering a question raised by [Khot05]. Previous hardness results for SVP, and the related minimum distance problem (MDP) for error-correcting codes, all use lattice/coding gadgets that have an abundance of codewords in a ball of radius smaller than the minimum distance. In contrast, our reduction only needs many codewords in a ball of radius slightly larger than the minimum distance. This enables an easy derandomization of our reduction for finite fields, giving a new elementary proof of deterministic hardness for MDP. We believe this weaker density requirement might offer a promising approach to showing deterministic hardness of SVP, a long elusive goal. The key technical ingredient underlying our result for real subspaces is a proof that in the kernel of a random Rademacher matrix, the support of any two linearly independent vectors have very little overlap. A broader motivation behind this work is the development of inapproximability techniques for problems over the reals. We hope that the approach we develop could enable progress on analytic variants of sparsest vector.